Lightweight CNN for Sorghum Plant Counting: A Patch-Wise Approach

Agriculture plays a vital role in the economies of developing nations, serving as both a subsistence and a source of wealth. To mitigate the risks of insufficient production and food scarcity, numerous agricultural output forecasting techniques have been developed. While traditional methods often rely on satellite data, recent advancements in deep learning and UAV (Unmmaned Aerial Vehicle) technology have opened new avenues for plant phenotyping. Early-stage plant counting is a key strategy for yield forecasting. However, manual crop counting is a time-consuming and labor-intensive task. To address this, high-throughput, automated plant counting methods are essential. This study proposes an approach for plant counting in sorghum crops using aerial images. Our method employs a patch-wise approach, where each image is divided into smaller patches. A complexity-optimized DCNN model is then applied to each patch to estimate the number of plants. By utilizing a regression layer, the model predicts the plant count for each patch. Our proposed method outperforms solutions based on known CNN architectures on the same dataset, achieving a lower Mean Absolute Error (MAE) with fewer parameters. This demonstrates the effectiveness of our approach in accurately counting plants in agricultural fields.